betaFeedback

WorksComputer Science

Federated Physics-Informed Thermal Field Prediction Across Unseen Geometries in Distributed Additive Manufacturing

Abstract

In additive manufacturing (AM), each change in part geometry requires a new thermal simulation, which makes design evaluation over many candidates costly. Geometry-aware models reduce this cost by predicting thermal fields on unseen parts with a single learned model, but they assume that diverse geometries are available centrally. This assumption is restrictive when each manufacturing site holds only a limited set of proprietary part designs that it cannot share. We propose Fed-PIGANO, a federated framework that trains physics-informed geometry-aware neural operator (PI-GANO) models at individual sites and aggregates their geometry encoders and prediction networks without sharing part geometries. We evaluate the framework on a 2D steady-state thermal benchmark for metal AM, with geometries from five polygonal families distributed across three sites under four scenarios that range from identical to strongly different family compositions. Over seven replications, Fed-PIGANO achieves mean relative L2 errors of 1.83% to 2.04% across all 12 scenario-site cases, compared with up to 20.46% for local-only training, and consistently yields lower PDE residuals. Its prediction accuracy is comparable to that of a centralized model trained on the pooled geometries of all sites. These results demonstrate that federated learning lets manufacturing sites benefit from each other's geometries without sharing them, enabling rapid thermal prediction on new part designs for tasks such as shape optimization, inverse design, and uncertainty quantification.

Introduction

Many physical phenomena in engineering, including heat transfer, fluid flow, and structural deformation, are described by systems of partial differential equations (PDEs). Numerical methods such as the finite element method (FEM) approximate their solutions by discretizing the geometric domain into a mesh [9]. For systems with multiscale and multiphysics behavior, however, such solvers become computationally demanding [31, 30]. This cost is particularly limiting in many-query tasks such as design optimization, which may require a large number of simulations [4, 19].

The design phase of thermal-based metal additive manufacturing (AM) is a representative many-query problem, where candidate part geometries are repeatedly evaluated and revised [29, 1]. In these processes, a localized heat source melts material that then solidifies, building up the part layer by layer [35]. The source may be stationary, as in spot melting, where a laser exposes a single location [7, 34, 39]. It may also move across a powder bed, as in laser powder bed fusion (L-PBF) [35]. In both cases, the heat diffusion PDE governs the thermal behavior over the part’s specific geometry. It couples material properties (e.g., thermal conductivity, density, specific heat), process parameters (e.g., heat source power, beam size, scan speed), and boundary interactions with the surroundings [3]. The resulting thermal history of the part is central to its process–structure–property (PSP) relationship. It drives microstructure evolution, which in turn determines the material’s mechanical properties and, ultimately, the manufactured part’s quality [31, 2]. A representative thermal-based PSP framework for a fixed geometry is presented in our earlier work (see Figure 1 of Saimon et al., [28]).

Distributed-data challenge in AM

Figure 1. Distributed-data challenge in AM. (a) Each of the CC sites holds a limited set of ncn_c part geometries. (b) Local-only training keeps data private but covers few geometries. (c) Pooling broadens coverage but may be restricted by data-sharing constraints.

Under fixed process parameters, any change to the part shape changes the domain boundary and therefore the thermal field. Such changes are common during the design phase of AM, for example to improve performance, reduce material usage, or meet new requirements [29]. Design tasks such as shape optimization, inverse design, uncertainty quantification, and design exploration follow an iterative loop around these changes. In iteration ii, the thermal field T(x,y;Ω(i))T(x, y; \Omega^{(i)}) of a candidate geometry Ω(i)\Omega^{(i)} is evaluated, and a performance metric is computed from it (e.g., peak temperature, cooling rate, or a microstructure-related quality index). The geometry is then updated to Ω(i+1)\Omega^{(i+1)}, for example by a gradient-based or heuristic search step. Each update changes the thermal field, so the structure and property predictions computed from the previous field no longer apply. In the conventional workflow, the new geometry must then be meshed and solved with FEM again. As iterations accumulate, this repeated cost substantially lengthens the part development cycle [29].

Physics-informed neural networks (PINNs) [27] incorporate the governing PDE and boundary conditions directly into model training, which reduces the dependence on labeled thermal solutions. Conventional PINN formulations for AM, however, consider a fixed geometric domain, so a change in part geometry typically requires retraining the model [28, 3]. Geometry-aware physics-informed models remove this limitation by taking the domain geometry as an input to the prediction model. Once trained on a set of known geometries {Ω(1),…,Ω(n)}\{\Omega^{(1)}, \ldots, \Omega^{(n)}\}, such a model with trainable parameters θ\theta predicts the thermal field Tθ(x,y;Ωnew)T_{\theta}(x,y;\Omega^{\mathrm{new}}) on an unseen geometry Ωnew\Omega^{\mathrm{new}} without retraining or a new FEM solve [40, 28]. Each design iteration then needs only a model evaluation, and the microstructure, property, and quality predictions that build on the thermal field are obtained faster as well [28].

The ability of such models to generalize to unseen part shapes depends on the geometric diversity of the known geometries used for training [40, 28]. In distributed AM, however, design and production take place independently at multiple manufacturing sites, and each site accumulates its own geometry dataset, shaped by the parts it has produced (Figure 1a). A site cc may therefore hold only ncn_c geometries, which cover only part of the geometric variation across the manufacturing network.

This distributed setting introduces two practical constraints for training geometry-aware models. First, a model trained only on a site’s local dataset keeps its geometries private but sees limited geometric diversity (Figure 1b), so it may predict poorly on geometries that are rare or absent locally. Second, pooling the datasets of multiple sites would broaden geometric coverage (Figure 1c), but such data sharing may be restricted by proprietary design information, intellectual property considerations, contractual agreements, or regulatory requirements. Each site could instead add synthetic geometries, but these may not reflect the part designs produced at other sites.

These constraints motivate the use of federated learning (FL) [23, 10] (Li et al. 2020), which enables multiple manufacturing sites to collaboratively train a shared model while keeping their part geometries local. Each site trains the model using its own local data and communicates only model parameters to a coordinating server, where the updates are aggregated into a shared global model over successive communication rounds. In our setting, all sites share the same metal AM thermal process, with the same steady-state heat conduction equation, Gaussian laser source, and material properties. Under this setting, site-to-site differences arise from differences in the local distributions of part geometries rather than from differences in the underlying thermal physics or process conditions (Figure 1a). Federated training must therefore aggregate both the geometry encoder and the prediction network of the local models. The resulting global model learns from the geometry families of all sites and can predict on unseen geometries beyond any single site’s training set, without pooling the geometries.

In this paper, we make the following contributions:

  • We formulate federated thermal-field prediction across unseen AM part geometries, where manufacturing sites collaboratively train a geometry-aware model without pooling proprietary part designs.

  • We propose Fed-PIGANO, a federated physics-informed geometry-aware framework that adopts PI-GANO as the local model at each site, trains it with a physics-informed loss without labeled temperature data, and aggregates its geometry encoder and pre-

diction network across sites.

  • We evaluate Fed-PIGANO on 2D metal AM geometries from five families, held by three sites under four scenarios with increasingly different family compositions, over seven replications. Its mean relative error on unseen geometries stays between 1.8% and 2.0% at every site, whereas that of training at each site alone (Local-only) reaches 20.5%. Its accuracy is also comparable to that of a centralized model trained on the pooled geometries.

The remainder of this paper is organized as follows. § 2 reviews three lines of work that converge in our approach. § 3 formulates the problem and presents Fed-PIGANO. § 4 describes the experimental setup, followed by the results and discussion in § 5. Finally, § 6 concludes the paper and outlines directions for future extension.

Related Work

This section reviews physics-informed thermal modeling in AM (§ 2.1), geometry-aware physics-informed learning (§ 2.2), and federated learning in manufacturing and scientific machine learning (§ 2.3). § 2.4 then summarizes the research gap at their intersection.

Physics-Informed Thermal Modeling in Additive Manufacturing

Physics-informed learning [27, 6] has been increasingly explored for thermal modeling in AM [41, 37, 19, 5]. A broader review of PINN applications in AM is provided by Faegh et al. [3]. These AM studies primarily consider a fixed part geometry during training and inference, referred to as the single-geometry PINN setting by Saimon et al. [28], Figure 2, top. In this setting, the model predicts at unseen coordinates within the same geometric domain, but a change in part geometry requires retraining the model. This limits its use in design tasks that repeatedly evaluate different geometries [13, 29, 40]. Geometry-aware physics-informed models, reviewed in § 2.2, address this limitation by learning across multiple geometries.

Fed-PIGANO framework

Figure 2. Fed-PIGANO framework. Each site c∈{1,…,C}c \in\{1,\ldots,C\} trains a local model θc\theta_c that maps Xb\mathbf{X}_b and (x,y)(x,y) to TθT_{\theta}, using only its own geometries. In each round, the server aggregates all model components into θr=∑cwcθcr\theta^r = \sum_c w_c \theta_c^r and broadcasts it back. The trained global model predicts thermal fields on unseen geometries.

Geometry-Aware Physics-Informed Learning

PIPN [13] introduced geometry-aware physics-informed learning by combining PINNs [27] with the point-cloud-based PointNet architecture (Qi et al., 2017). PIPN and its extensions have been applied to fluid flow, structural mechanics, porous-media transport, aerodynamics, vascular flow, crack propagation, and acoustic–structure interaction [13, 12, 40, 14, 21, 11, 33] (Tu et al. 2022). More recently, PI-KAN-PointNet [15] replaced the shared MLPs of the PIPN architecture with Kolmogorov–Arnold Networks (KANs) [22]. GeoPIPN [28] extends geometry-aware physics-informed prediction to thermal fields on unseen AM part geometries. PI-GANO [40], in contrast, is a general physics-informed geometry-aware neural operator that handles variations in both geometry and PDE parameters. Generalization in scientific machine learning may thus concern the geometric domain, the PDE parameters, or both. This work focuses on geometry generalization with fixed material and process parameters, and we adopt the geometry-only variant of PI-GANO as the local model at each site (§ 3.4). All of these approaches, however, train on a centralized dataset and assume that all training geometries are accessible in one place. This assumption does not hold when the geometries are held by different AM sites.

Federated Learning in Manufacturing and Scientific Machine Learning

FL has been explored in manufacturing, particularly in AM, to enable collaborative modeling while keeping the data at local sites. Mehta et al. (2024) applied FL to dimensional accuracy prediction and part qualification across AM factories. For metal L-PBF, Mehta and Shao (2022) considered federated defect detection, while Lei et al. (2024) investigated federated heat-emission prediction across small and medium-sized enterprises. Collaborative learning has also been investigated for process-parameter recommendation across networked 3D printers [8]. These studies show that collaborative learning can benefit distributed manufacturing, but they do not consider physics-informed thermal field prediction across varying AM part geometries.

In the broader scientific machine learning literature, Zhang et al. [38] presented a systematic study of federated learning for function approximation and PDE solving. They introduced FedAvg-Adam, a training strategy that replaces the stochastic gradient descent of FedAvg [23] with Adam for local updates. In their federated PINN formulation, the collocation points of a single square domain are divided among clients. All clients thus solve the same PDE on the same domain, and the samples distributed across clients are spatial coordinates rather than different domains. This formulation therefore does not require a geometry encoder to represent different domain shapes across clients, and the study does not address AM thermal problems.

Research Gap

Table 1 summarizes the three lines of work reviewed above. Physics-informed thermal models for AM (§ 2.1) mostly consider a single fixed geometry. Geometry-aware physics-informed models (§ 2.2) predict across varying domains but assume that all training geometries are available in one place. Federated learning studies in manufacturing and scientific machine learning (§ 2.3) keep data local but do not consider varying geometric domains across clients. To our knowledge, no existing work combines physics-informed learning, geometry generalization, and federated training for AM thermal-field prediction. This work addresses this gap with Fed-PIGANO, which trains a geometry-aware physics-informed model across AM sites without pooling their part geometries.

StudyPhysics-informedGeometry-awareFederated learningAM thermal focus
PIPN [13]✓✓××
PI-KAN-PointNet [15]✓✓××
PI-GANO [40]✓✓××
coPINN [5]✓××✓
GeoPIPN [28]✓✓×✓
FedAvg-Adam [38]✓×✓×
This work✓✓✓✓

Table 1. Comparison of related studies and the research gap addressed in this work.

Methodology

This section first formulates thermal prediction on unseen geometries across AM sites (§ 3.1) and the thermal model that all sites share (§ 3.2). It then presents the Fed-PIGANO framework (§ 3.3), followed by its local geometry-aware model (Section § 3.4 and physics-informed loss (§ 3.5). In this paper, we consider 2D AM part geometries, so all equations in this section are written in the in-plane coordinates (x,y)(x, y).

Problem Formulation for Cross-Site Thermal Prediction on Unseen Geometries

We consider CC clients, indexed by c∈{1,…,C}c \in\{1, \ldots, C\}, each representing an AM site that designs and builds parts independently. All clients share the thermal problem of § 3.2. They differ only in the part geometries they hold, which reflect each site’s own design history and may cover only a subset of the geometric variation across the network. In our experiments, this site-to-site difference takes the form of different geometry-family compositions (§ 4.2). Client cc holds a training set Dctr\mathcal{D}^{\mathrm{tr}}_c of ncn_c geometries, a validation set Dcval\mathcal{D}^{\mathrm{val}}_c, and a test set Dcte\mathcal{D}^{\mathrm{te}}_c. Each geometry gg is a domain Ωg⊂R2\Omega^g \subset\mathbb{R}^2 that represents a 2D AM part (§ 3.2.1). It is discretized by a point set Xg=Xbg∪Xintg\mathbf{X}^g = \mathbf{X}^g_b \cup\mathbf{X}^g_{\mathrm{int}}, with nbn_b boundary points Xbg⊂∂Ωg\mathbf{X}^g_b \subset\partial\Omega^g and nintn_{\mathrm{int}} interior points Xintg⊂Ωg\mathbf{X}^g_{\mathrm{int}} \subset\Omega^g, so that Npts=nb+nintN_{\mathrm{pts}} = n_b + n_{\mathrm{int}} is the same for all geometries. The boundary point set Xbg\mathbf{X}^g_b describes the shape of the geometry, and the full point set Xg\mathbf{X}^g provides the collocation points for training. All coordinates are in the nondimensional form of § 3.2.2.

The goal is a single geometry-aware model Tθ(x,y;Xb)T_\theta(x, y; \mathbf{X}_b) that predicts the temperature at any point (x,y)(x, y) of a geometry Ω\Omega described by its boundary point set Xb\mathbf{X}_b, including new candidate geometries that no client has seen, such as those that arise when part designs are revised (§ 1). The geometries remain at their clients, and only model parameters are exchanged with a central server. The federated objective is

min⁡θJ(θ)=∑c=1Cwc L(θ;Dctr),wc=nc∑c′=1Cnc′(1)\min_{\theta}\mathcal{J}(\theta)=\sum_{c=1}^{C}w_c\,\mathcal{L}(\theta;\mathcal{D}^{\mathrm{tr}}_c), \qquad w_c=\frac{n_c}{\sum_{c'=1}^{C}n_{c'}} \tag*{(1)}

where L\mathcal{L} is the physics-informed loss of § 3.5, and the weights wcw_c follow FedAvg [23]. When all clients hold the same number of geometries, as in our experiments, wc=1/Cw_c=1/C.

AM Thermal Model

This subsection defines the steady-state heat conduction problem that all clients share and the nondimensional form used throughout the rest of the paper.

Steady-State Heat Conduction

We consider steady-state heat conduction on a geometry Ω\Omega, with the superscript gg dropped. The domain represents a thin horizontal cross-section of the part, i.e., an effective deposited layer of thickness hh, heated by a Gaussian laser source. The material properties, laser source, and boundary condition are the same for every geometry and every client, so the part geometry is the only quantity that changes the thermal field. The governing equation is

−k(∂2T(x,y)∂x2+∂2T(x,y)∂y2)=q(x,y),(x,y)∈Ω(2)-k\left(\frac{\partial^{2}T(x,y)}{\partial x^{2}}+\frac{\partial^{2}T(x,y)}{\partial y^{2}}\right)=q(x,y), \qquad(x,y)\in\Omega \tag*{(2)}

subject to the Dirichlet boundary condition

T(x,y)=Ta,(x,y)∈∂Ω(3)T(x,y)=T_a, \qquad(x,y)\in\partial\Omega \tag*{(3)}

where kk is the thermal conductivity, TT is the temperature, TaT_a is the ambient temperature, and qq is the volumetric heat source. The laser source is modeled as a circular Gaussian beam with absorbed power Qtotal=αPQ_{\mathrm{total}}=\alpha P, where PP is the laser power and α\alpha is the absorptivity:

q(x,y)=2Qtotalπrb2hexp⁡(−2[(x−xL)2+(y−yL)2]rb2)(4)q(x,y)=\frac{2Q_{\mathrm{total}}}{\pi r_b^{2}h}\exp\left(-\frac{2\left[(x-x_L)^2+(y-y_L)^2\right]}{r_b^{2}}\right) \tag*{(4)}

where rbr_b is the beam radius and (xL,yL)(x_L,y_L) is the beam center. Its prefactor makes the integral of the source over the layer, h∫R2q(x,y) dx dyh\int_{\mathbb{R}^{2}}q(x,y)\,\mathrm{d}x\,\mathrm{d}y, equal to QtotalQ_{\mathrm{total}}, so the absorbed power is distributed as a Gaussian in the plane and uniformly through the layer thickness.

The source is stationary, as in spot melting, where the laser exposes a single location ([7, 34, 39]). We assume that the exposure lasts long enough for the temperature to reach a steady state, and solve Eq. (2) as a boundary value problem with the source held at a fixed position (xL,yL)(x_L, y_L). The resulting field approximates the temperature during one exposure. Table 2 lists the source position together with all material and process parameters.

ParameterSymbolValueParameterSymbolValue
Thermal conductivitykk9.8 W m−1 K−19.8\ \mathrm{W\,m^{-1}\,K^{-1}}Laser powerPP375 W375\ \mathrm{W}
Ambient temperatureTaT_a300 K300\ \mathrm{K}Absorptivityα\alpha0.40.4
Effective layer thicknesshh2.5 mm2.5\ \mathrm{mm}Absorbed powerQtotal=αPQ_{\mathrm{total}}=\alpha P150 W150\ \mathrm{W}
Beam radiusrbr_b2.5 mm2.5\ \mathrm{mm}Source center(xL,yL)(x_L,y_L)(−1.6 rb, 0)(-1.6\,r_b,\,0)

Table 2. Material and process parameters of the thermal model.

Nondimensionalization

The parameters in Table 2 span several orders of magnitude. We therefore nondimensionalize the problem, which combines them into a single source amplitude and brings the network inputs, the output, and the PDE terms to order one. This matches the scale assumed by standard weight initialization and keeps the boundary and PDE loss terms comparable in magnitude, which improves the conditioning of training ([36]). We use the beam radius rbr_b as the length scale and the ambient temperature TaT_a as the temperature scale:

x=rbx∗,y=rby∗,T=TaT∗(5)x = r_b x^*, \qquad y = r_b y^*, \qquad T = T_a T^* \tag*{(5)}

With these scales, ∂2/∂x2=rb−2 ∂2/∂x∗2\partial^2/\partial x^2 = r_b^{-2}\,\partial^2/\partial x^{*2} and ∂2/∂y2=rb−2 ∂2/∂y∗2\partial^2/\partial y^2 = r_b^{-2}\,\partial^2/\partial y^{*2}, and the exponent of Eq. (4) becomes −2[(x∗−xL∗)2+(y∗−yL∗)2]-2\left[(x^*-x_L^*)^2+(y^*-y_L^*)^2\right] with xL∗=xL/rbx_L^*=x_L/r_b and yL∗=yL/rby_L^*=y_L/r_b. Substituting Eq. (5) into Eqs. (2)–(4) and dividing by kTa/rb2kT_a/r_b^2 gives

−(∂2T∗(x∗,y∗)∂x∗2+∂2T∗(x∗,y∗)∂y∗2)=q∗(x∗,y∗),(x∗,y∗)∈Ω∗(6)-\left(\frac{\partial^2 T^*(x^*,y^*)}{\partial x^{*2}}+\frac{\partial^2 T^*(x^*,y^*)}{\partial y^{*2}}\right)=q^*(x^*,y^*), \qquad(x^*,y^*)\in\Omega^* \tag*{(6)}
T∗(x∗,y∗)=1,(x∗,y∗)∈∂Ω∗(7)T^*(x^*,y^*)=1, \qquad(x^*,y^*)\in\partial\Omega^* \tag*{(7)}
q∗(x∗,y∗)=Λexp⁡(−2[(x∗−xL∗)2+(y∗−yL∗)2])(8)q^*(x^*,y^*)=\Lambda\exp\left(-2\left[(x^*-x_L^*)^2+(y^*-y_L^*)^2\right]\right) \tag*{(8)}

where Λ=2Qtotal/(πkTah)\Lambda= 2Q_{\mathrm{total}}/(\pi k T_a h) is the dimensionless source amplitude. With the values in Table 2, Λ≈12.99\Lambda\approx12.99 and (xL∗,yL∗)=(−1.6,0)(x_L^*, y_L^*) = (-1.6, 0). Eqs. (6)–(8) thus define the thermal problem shared by all clients, on every geometry considered in this study (§ 4.1). From here on, all quantities are nondimensional, and the asterisks are dropped for brevity.

Fed-PIGANO: Federated Training of the Geometry-Aware Thermal Model with FedAvg-Adam

Figure 2 illustrates the proposed Fed-PIGANO framework for the federated problem of Eq. (1). Each client trains a geometry-aware local model (§ 3.4) with a physics-informed loss (§ 3.5) that uses only the governing equation and boundary condition of § 3.2. In each communication round, the server broadcasts the global parameters, each client trains on its own geometries, and the server aggregates and validates the updated parameters. All components of the local model, including the geometry encoder, are aggregated, so the global model learns from the geometries of all clients. Instead of the stochastic gradient descent of standard FedAvg ([23]), we follow Zhang et al. [38] and use Adam ([16]) for local updates, keeping each client’s moment estimates mc\mathbf{m}_c and vc\mathbf{v}_c across rounds (FedAvg-Adam). Algorithm 1 summarizes the procedure, and the three steps of a round are described below.

Broadcast and local training. At the start of round rr, each client sets θc←θr−1\theta_c \leftarrow\theta^{r-1}, while mc\mathbf{m}_c and vc\mathbf{v}_c carry over from the previous round. The client then runs EE local epochs over Dctr\mathcal{D}_c^{\mathrm{tr}}. Let tt be the step counter of client cc, which is not reset between rounds. At step tt, the client takes a mini-batch B⊂Dctr\mathcal{B} \subset\mathcal{D}_c^{\mathrm{tr}} of at most BB geometries, computes ∇ct=∇θL(θc;B)\nabla_c^t = \nabla_\theta\mathcal{L}(\theta_c; \mathcal{B}) from Eq. (19), and updates

mct=β1mct−1+(1−β1)∇ct,m^ct=mct/(1−β1t),vct=β2vct−1+(1−β2)∇ct⊙∇ct,v^ct=vct/(1−β2t),θc←θc−ηm^ct⊘(v^ct+ϵ)(9)\begin{aligned} \mathbf{m}_c^t &= \beta_1 \mathbf{m}_c^{t-1} + (1-\beta_1)\nabla_c^t, \qquad&\widehat{\mathbf{m}}_c^t &= \mathbf{m}_c^t/(1-\beta_1^t),\\ \mathbf{v}_c^t &= \beta_2 \mathbf{v}_c^{t-1} + (1-\beta_2)\nabla_c^t \odot\nabla_c^t, \qquad&\widehat{\mathbf{v}}_c^t &= \mathbf{v}_c^t/(1-\beta_2^t),\\ \theta_c &\leftarrow\theta_c-\eta\widehat{\mathbf{m}}_c^t \oslash\left(\sqrt{\widehat{\mathbf{v}}_c^t}+\epsilon\right) \tag*{(9)} \end{aligned}

where ⊙\odot, ⊘\oslash, and the square root act elementwise, η\eta is the learning rate, and β1\beta_1, β2\beta_2, ϵ\epsilon are the Adam constants. The moments are not reset, so the adaptive step size of each client reflects its full gradient history rather than only the current round. After EE epochs, the client holds the updated parameters θcr\theta_c^r.

Aggregation. All clients participate in every round. The server aggregates their parameters as

θr=∑c=1Cwcθcr(10)\theta^r = \sum_{c=1}^{C} w_c \theta_c^r \tag*{(10)}

The global model is thus a parameter average of local models trained on different sets of geometries. Only model parameters are communicated, and the moment estimates mc\mathbf{m}_c and vc\mathbf{v}_c stay at each client.

Validation and model selection. After aggregation, each client evaluates θr\theta^r on its validation set and returns the scalar loss to the server, which forms

Lvalr=∑c=1CwcL(θr;Dcval)(11)\mathcal{L}_{\mathrm{val}}^r = \sum_{c=1}^{C} w_c \mathcal{L}\left(\theta^r;\mathcal{D}_c^{\mathrm{val}}\right) \tag*{(11)}

The server uses Lvalr\mathcal{L}_{\mathrm{val}}^r for two purposes. First, it controls a single learning rate η\eta shared by all clients. The rate starts at η0\eta_0 and is reduced by a fixed factor, down to a floor ηmin⁡\eta_{\min}, with ReduceLROnPlateau when Lvalr\mathcal{L}_{\mathrm{val}}^r has not decreased by a relative threshold for a set number of rounds (the patience). After each reduction, the rate is held fixed for a cooldown period before it can be reduced again. Second, it selects the global checkpoint with the lowest

Algorithm 1 Fed-PIGANO training with FedAvg-Adam

Require: Client sets {Dctr,Dcval}c=1C\{\mathcal{D}_c^{\mathrm{tr}},\mathcal{D}_c^{\mathrm{val}}\}_{c=1}^{C}; weights wcw_c; rounds RR; local epochs EE; batch size BB; η0,β1,β2,ε\eta_0,\beta_1,\beta_2,\varepsilon; loss weights λbc,λpde\lambda_{\mathrm{bc}},\lambda_{\mathrm{pde}}; scheduler settings (Table 4).

SettingLocal-onlyFed-PIGANOCentralized
Training / validation geometries50 / 2550 / 25 per client150 / 75 (pooled)
Hidden width dhd_h128
Weight initializationGlorot uniform (tanh gain), zero biases
OptimizerAdam (η0=10−3\eta_0 = 10^{-3}, β1=0.9\beta_1 = 0.9, β2=0.999\beta_2 = 0.999, ε=10−8\varepsilon= 10^{-8})
Learning-rate scheduleReduceLROnPlateau (factor 0.5, relative threshold 5×10−35 \times10^{-3})
Patience / cooldown500 / 100 epochs100 / 20 rounds500 / 100 epochs
Learning-rate floor ηmin⁡\eta_{\min}10−610^{-6}
Mini-batch size BB (geometries)8
Loss weightsλbc=λpde=1\lambda_{\mathrm{bc}} = \lambda_{\mathrm{pde}} = 1
Training lengthEmax⁡=12,000E_{\max} = 12{,}000 epochsR=2,400R = 2{,}400 rounds, E=5E = 5Emax⁡=12,000E_{\max} = 12{,}000 epochs
Model selectionEq. (21)Eq. (12)Eq. (21)

Table 4. Training settings of the three models. Shared settings span all three columns.

  1. Initialize θ0=(θgeo,θtr,θdec)\boldsymbol{\theta}^{0}=(\boldsymbol{\theta}_{\mathrm{geo}},\boldsymbol{\theta}_{\mathrm{tr}},\boldsymbol{\theta}_{\mathrm{dec}}) ▷ § 4.4

  2. At every client: mc,vc←0\mathbf{m}_c,\mathbf{v}_c\leftarrow0, t←0t\leftarrow0; at the server: η←η0\eta\leftarrow\eta_0, Lval∗←+∞\mathcal{L}_{\mathrm{val}}^{*}\leftarrow+\infty

  3. for r=1,…,Rr=1,\ldots,R do

  4.     for each client c=1,…,Cc=1,\ldots,C in parallel do

  5.         θc←θr−1\boldsymbol{\theta}_c\leftarrow\boldsymbol{\theta}^{r-1} ▷ Broadcast θr−1\boldsymbol{\theta}^{r-1} and η\eta; mc,vc\mathbf{m}_c,\mathbf{v}_c kept

  6.         for e=1,…,Ee=1,\ldots,E do

  7.             for each mini-batch B⊂Dctr\mathcal{B}\subset\mathcal{D}_c^{\mathrm{tr}}, ∣B∣≤B|\mathcal{B}|\le B do ▷ random partition

  8.                 t←t+1t\leftarrow t+1

  9.                 ∇ct←∇θL(θc;B)\nabla_c^t\leftarrow\nabla_{\boldsymbol{\theta}}\mathcal{L}(\boldsymbol{\theta}_c;\mathcal{B}) ▷ Eq. (19)

  10.                 Update mc,vc,θc\mathbf{m}_c,\mathbf{v}_c,\boldsymbol{\theta}_c ▷ Eq. (9)

  11.             end for

  12.         end for

  13.         Send θcr←θc\boldsymbol{\theta}_c^r\leftarrow\boldsymbol{\theta}_c to the server

  14.     end for

  15.     θr←∑cwcθcr\boldsymbol{\theta}^r\leftarrow\sum_c w_c\boldsymbol{\theta}_c^r ▷ Aggregation, Eq. (10)

  16.     Each client returns L(θr;Dcval)\mathcal{L}(\boldsymbol{\theta}^r;\mathcal{D}_c^{\mathrm{val}}) ▷ Validation and model selection

  17.     Lvalr←∑cwcL(θr;Dcval)\mathcal{L}_{\mathrm{val}}^r\leftarrow\sum_c w_c\mathcal{L}(\boldsymbol{\theta}^r;\mathcal{D}_c^{\mathrm{val}}) ▷ Eq. (11)

  18. &nbsp;&nbsp;&nbsp;&nbsp;if Lvalr<Lval∗\mathcal{L}_{\mathrm{val}}^r<\mathcal{L}_{\mathrm{val}}^{*} then

  19. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;θ∗←θr\boldsymbol{\theta}^{*}\leftarrow\boldsymbol{\theta}^r; r∗←rr^{*}\leftarrow r; Lval∗←Lvalr\mathcal{L}_{\mathrm{val}}^{*}\leftarrow\mathcal{L}_{\mathrm{val}}^r ▷ Eq. (12)

  20. &nbsp;&nbsp;&nbsp;&nbsp;end if

  21. &nbsp;&nbsp;&nbsp;&nbsp;η←ReduceLROnPlateau⁡(η,Lvalr)\eta\leftarrow\operatorname{ReduceLROnPlateau}(\eta,\mathcal{L}_{\mathrm{val}}^r) ▷ Shared by all clients

  22. end for

  23. return θ∗\boldsymbol{\theta}^{*}

validation loss,

r∗=arg⁡min⁡r∈{1,…,R}Lvalr,θ∗=θr∗(12)r^{*}=\arg\min_{r\in\{1,\ldots,R\}}\mathcal{L}_{\mathrm{val}}^r,\qquad\boldsymbol{\theta}^{*}=\boldsymbol{\theta}^{r^{*}} \tag*{(12)}

Model selection thus requires no reference solutions, as L\mathcal{L} uses only the governing equation and boundary condition (§ 3.5). The values of all training settings are given in § 4.4.

Local Geometry-Aware Model (PI-GANO)

We adopt the geometry-only variant of PI-GANO [40] as the local model of Fed-PIGANO at each client, since it predicts PDE solutions on varying domains given the geometry. Hereafter, we refer to this variant as PI-GANO. The model has three trainable components, shown for each client in Figure 2, a geometry encoder with parameters θgeo\theta_{\mathrm{geo}}, a trunk network with parameters θtr\theta_{\mathrm{tr}}, and a decoder with parameters θdec\theta_{\mathrm{dec}}. We write θ=(θgeo,θtr,θdec)\theta= (\theta_{\mathrm{geo}}, \theta_{\mathrm{tr}}, \theta_{\mathrm{dec}}) for the full set of weights and biases.

Following Zhong and Meidani [40], the geometry encoder Fgeo\mathcal{F}_{\mathrm{geo}} is a four-layer fully connected network with widths 2→dh→dh→dh→dh2 \to d_h \to d_h \to d_h \to d_h, tanh activations on the hidden layers, and a linear output layer. It is applied with shared weights to every boundary point, and the resulting features are averaged into the geometry embedding

zΩ=1nb∑(x,y)∈XbFgeo([x,y]T)∈Rdh(13)z_{\Omega} = \frac{1}{n_b} \sum_{(x,y)\in X_b} \mathcal{F}_{\mathrm{geo}}\left([x,y]^{\mathsf{T}}\right) \in\mathbb{R}^{d_h} \tag*{(13)}

Mean pooling makes zΩz_{\Omega} independent of the ordering of the boundary points [28] (Qi et al. 2017), and the same embedding is used for every query point on that geometry. The trunk network is a single linear layer that lifts the query point to u(x,y)=Wtr[x,y]T+btr∈Rdh\mathbf{u}(x,y) = \mathbf{W}^{\mathrm{tr}}[x,y]^{\mathsf{T}} + \mathbf{b}^{\mathrm{tr}} \in\mathbb{R}^{d_h}. The two features are then concatenated and passed to the decoder Fdec\mathcal{F}_{\mathrm{dec}}, a fully connected network of the same form with widths 2dh→2dh→2dh→2dh→12d_h \to2d_h \to2d_h \to2d_h \to1:

Tθ(x,y;Xb)=Fdec([u(x,y)∥zΩ])(14)T_{\theta}(x,y;X_b) = \mathcal{F}_{\mathrm{dec}}\left([\mathbf{u}(x,y) \mathbin{\|} z_{\Omega}]\right) \tag*{(14)}

where ∥\mathbin{\|} denotes concatenation. The prediction therefore depends on both the query location and the shape of the domain. All three components are trained locally with the physics-informed loss of § 3.5 and aggregated at the server by Eq. (10).

Physics-Informed Local Loss Formulation

We train PI-GANO with a physics-informed loss derived only from the governing equation and boundary condition of § 3.2, without temperature labels from either simulations or sensors. FEM solutions serve only as reference fields for evaluation (§ 4.5). For a single geometry gg, the loss combines a boundary-condition term and a PDE-residual term,

ℓ(θ;g)=λbcℓbc(θ;g)+λpdeℓpde(θ;g)(15)\ell(\theta;g) = \lambda_{\mathrm{bc}} \ell_{\mathrm{bc}}(\theta;g) + \lambda_{\mathrm{pde}} \ell_{\mathrm{pde}}(\theta;g) \tag*{(15)}

where λbc\lambda_{\mathrm{bc}} and λpde\lambda_{\mathrm{pde}} are loss weights, both set to 1 (Table 4). The boundary term enforces the Dirichlet condition (7) on the boundary point set:

ℓbc(θ;g)=1nb∑(x,y)∈Xbg[Tθ(x,y;Xbg)−1]2(16)\ell_{\mathrm{bc}}(\theta;g)=\frac{1}{n_b}\sum_{(x,y)\in X_b^g}\left[T_\theta(x,y;X_b^g)-1\right]^2 \tag*{(16)}

The PDE term enforces Eq. (6) at every point of the full point set XgX^g, including the boundary points:

ℓpde(θ;g)=1Npts∑(x,y)∈Xg[Rθ(x,y;Xbg)]2(17)\ell_{\mathrm{pde}}(\theta;g)=\frac{1}{N_{\mathrm{pts}}}\sum_{(x,y)\in X^g}\left[R_\theta(x,y;X_b^g)\right]^2 \tag*{(17)}

where the pointwise residual is

Rθ(x,y;Xb)=−(∂2Tθ∂x2+∂2Tθ∂y2)−Λexp⁡(−2[(x−xL)2+(y−yL)2])(18)R_\theta(x,y;X_b)=-\left(\frac{\partial^2 T_\theta}{\partial x^2}+\frac{\partial^2 T_\theta}{\partial y^2}\right)-\Lambda\exp\left(-2\left[(x-x_L)^2+(y-y_L)^2\right]\right) \tag*{(18)}

with Λ\Lambda and (xL,yL)(x_L,y_L) as given in Section 3.2.2. The second derivatives are obtained by automatic differentiation of TθT_\theta with respect to xx and yy.

For any set SS of geometries, such as a mini-batch, a training set, or a validation set, we define the loss as the average over its members,

L(θ;S)=1∣S∣∑g∈Sℓ(θ;g)(19)\mathcal{L}(\theta;S)=\frac{1}{|S|}\sum_{g\in S}\ell(\theta;g) \tag*{(19)}

The residual (18) has the same operator and source term at every client, and client data enter the loss only through the geometry. All local models therefore learn the same process physics on different sets of geometries, which makes aggregating their parameters by Eq. (10) meaningful.

Experimental Setup

We evaluate Fed-PIGANO on the 2D AM thermal problem of Section 3.2. This section describes the geometry generation (§ 4.1), the client scenarios (§ 4.2), the Local-only baseline and the Centralized reference (§ 4.3), the implementation and the seven replications (§ 4.4), and the evaluation metrics (§ 4.5).

Random Geometry Generation and Discretization

Each domain Ω\Omega belongs to one of F=5F=5 polygonal geometry families. The Quadrilateral (Q), Pentagonal (P), and Hexagonal (H) families are convex, and the Concave-Pentagonal (CP) and Concave-Hexagonal (CH) families are concave. Figure 3 shows how we generate the geometries of each family. The vertices V1V_1 and V2V_2 are fixed where the domain centerline meets a reference circle of radius 10 mm centered at OO, so that every geometry has the same extent along the centerline and contains the fixed source position (§ 3.2.1). Each remaining vertex ViV_i is sampled uniformly within a circle of radius 3 mm around its nominal position. In the concave families, the sampling region of V5V_5 lies inside the reference circle, which creates a reentrant corner. Each family comprises 300 distinct geometries, which are randomly partitioned among the C=3C = 3 clients, so that each client holds a local pool of 100 geometries per family (500 in total).

Geometry generation scheme for the five geometry families

Figure 3. Geometry generation scheme for the five families. Orange polygons show one example geometry per family.

For each geometry, we solve the thermal problem of Eqs. (2)–(4) with the finite element method (FEM) using the MATLAB PDE solver, with a maximum element size of rb/28r_b/28. The point set X\mathbf{X} of each geometry (§ 3.1) is a random sample of nb=500n_b = 500 boundary nodes and nint=1,000n_{\mathrm{int}} = 1{,}000 interior nodes of the mesh, so that Npts=1,500N_{\mathrm{pts}} = 1{,}500. The mesh thus supplies the training point locations, while the FEM temperatures at these nodes serve only as reference values for evaluation (§ 4.5).

Client Geometry Distributions and Scenarios

To emulate manufacturing sites whose design geometries cover different parts of the geometry space (§ 3.1), we construct four scenarios in which the clients’ training geometries differ in family composition. From its local pool of 500 geometries (§ 4.1), each client draws disjoint training, validation, and test sets in each replication of the experiment (§ 4.4). The test set DcteD_c^{\mathrm{te}} contains 10 geometries per family (G=50G = 50). Within a replication, it is drawn once per client and reused in every scenario, so differences in test performance across scenarios reflect only the training and validation sets. The training set DctrD_c^{\mathrm{tr}} always contains nc=50n_c = 50 geometries, and the validation set DcvalD_c^{\mathrm{val}} contains 25 geometries with the same family proportions. The scenarios differ only in how the 50 training geometries are split across the five families. Table 3 lists the family counts of all sets.

We quantify the difference in family composition with the Jensen–Shannon divergence generalized to CC distributions ([20]). Let ncfn_c^f be the number of training geometries of family ff at client cc, and let pc=(pc1,…,pcF)\mathbf{p}_c = (p_c^1, \ldots, p_c^F), with pcf=ncf/ncp_c^f = n_c^f/n_c, be the family-proportion vector of client cc. Let pˉ=C−1∑c=1Cpc\bar{\mathbf{p}} = C^{-1}\sum_{c=1}^{C}\mathbf{p}_c be the global mixture. With equal client weights, this divergence is bounded above by ln⁡C\ln C ([20]), and we normalize it by this bound as

δ=1ln⁡C[H(pˉ)−1C∑c=1CH(pc)]∈[0,1](20)\delta= \frac{1}{\ln C}\left[H(\bar{\mathbf{p}})-\frac{1}{C}\sum_{c=1}^{C}H(\mathbf{p}_c)\right]\in[0,1] \tag*{(20)}

where H(p)=−∑f=1Fpfln⁡pfH(\mathbf{p})=-\sum_{f=1}^{F}p^f\ln p^f is the Shannon entropy, with 0ln⁡0=00\ln0=0. The value of δ\delta is 0 when all clients have the same family proportions and 1 when no two clients share a family. Here, δ\delta measures heterogeneity in family composition, not the similarity between individual shapes. In S1, every client has 10 training geometries per family. From S2 to S4, Client 1 trains increasingly on the convex families Q and P, Client 2 on H, and Client 3 on the concave families CP and CH. In S4, the only shared families are P (Clients 1 and 2) and CP (Clients 2 and 3). In every scenario, each family has 30 training geometries in total, so the global mixture pˉ\bar{p} is uniform and changes in δ\delta reflect only differences between clients. The divergence increases from δ=0\delta= 0 (S1) to 0.198 (S2), 0.449 (S3), and 0.768 (S4).

Local-Only Baseline and Centralized Reference

The Local-only baseline represents the setting of Figure 1b, in which each site trains a geometry-aware model on its own geometries only. Each client cc trains its own PI-GANO model θcloc\theta_c^{\mathrm{loc}} on Dctr\mathcal{D}_c^{\mathrm{tr}} without communication, using the same architecture (§ 3.4), loss (Eq. (19)), initialization, optimizer, mini-batch size, and learning-rate schedule as Fed-PIGANO. Training runs for Emax⁡E_{\max} epochs. After each epoch ee, the model θcloc,e\theta_c^{\mathrm{loc},e} is evaluated on Dcval\mathcal{D}_c^{\mathrm{val}}, and this validation loss both drives the learning-rate schedule and selects the checkpoint,

ec∗=arg⁡min⁡e∈{1,…,Emax⁡} L(θcloc,e;Dcval),θcloc=θcloc,ec∗(21)e_c^* = \underset{e \in\{1,\ldots,E_{\max}\}}{\arg\min}\ \mathcal{L}\left(\theta_c^{\mathrm{loc},e};\mathcal{D}_c^{\mathrm{val}}\right), \qquad\theta_c^{\mathrm{loc}} = \theta_c^{\mathrm{loc},e_c^*} \tag*{(21)}

which mirrors the checkpoint selection of Fed-PIGANO (Eq. (12)) at a single client. We set Emax⁡=R×EE_{\max} = R \times E, so each Local-only model makes the same number of passes over its training set as each client does in Fed-PIGANO. The scheduler settings in Table 4 are equivalent in the same sense, since a patience of 100 rounds and a cooldown of 20 rounds correspond to 500 and 100 epochs at E=5E = 5. We also train a Centralized reference, which represents the pooling setting of Figure 1c. It is a single PI-GANO model θcen\theta^{\mathrm{cen}} trained like the Local-only models, using the pooled training and validation sets of all clients. Centralized requires all geometries in one place, which our setting rules out, so it serves only as a reference, not as a competing method.

Implementation Details

Table 4 lists the training settings of the three models. Local-only and Fed-PIGANO differ only in federation (§ 4.3), so the performance differences between them reported in § 5 can be attributed to federated training. The PI-GANO architecture follows Zhong and Meidani [40] (§ 3.4). We repeat the experiment seven times with random seeds fixed in advance and report all seven replications. The seed of a replication sets the split of each client’s geometries into training, validation, and test sets with the counts of Table 3, the initial weights, and the mini-batch order. Within a replication, every Local-only model, the initial global model θ0\theta^0 of Fed-PIGANO, and the Centralized model start from identical parameters. We implemented all experiments in PyTorch 2.8.0 and ran them on a single NVIDIA GeForce RTX 4090 GPU. One training run takes about 13 min for a Local-only model, 36 min for Centralized, and 39 min for Fed-PIGANO, whose three clients are simulated one after another.

ScenarioSplitClient 1Client 2Client 3
S1 (δ=0.000\delta= 0.000)Train (50)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)
S1 (δ=0.000\delta= 0.000)Val (25)(5, 5, 5, 5, 5)(5, 5, 5, 5, 5)(5, 5, 5, 5, 5)
S1 (δ=0.000\delta= 0.000)Test (50)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)
S2 (δ=0.198\delta= 0.198)Train (50)(22, 16, 6, 4, 2)(4, 10, 18, 12, 6)(4, 4, 6, 14, 22)
S2 (δ=0.198\delta= 0.198)Val (25)(11, 8, 3, 2, 1)(2, 5, 9, 6, 3)(2, 2, 3, 7, 11)
S2 (δ=0.198\delta= 0.198)Test (50)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)
S3 (δ=0.449\delta= 0.449)Train (50)(26, 20, 4, 0, 0)(4, 8, 22, 12, 4)(0, 2, 4, 18, 26)
S3 (δ=0.449\delta= 0.449)Val (25)(13, 10, 2, 0, 0)(2, 4, 11, 6, 2)(0, 1, 2, 9, 13)
S3 (δ=0.449\delta= 0.449)Test (50)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)
S4 (δ=0.768\delta= 0.768)Train (50)(30, 20, 0, 0, 0)(0, 10, 30, 10, 0)(0, 0, 0, 20, 30)
S4 (δ=0.768\delta= 0.768)Val (25)(15, 10, 0, 0, 0)(0, 5, 15, 5, 0)(0, 0, 0, 10, 15)
S4 (δ=0.768\delta= 0.768)Test (50)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)(10, 10, 10, 10, 10)

Table 3. Per-client family counts (Q, P, H, CP, CH) of the training, validation, and test sets in each scenario. Larger δ\delta (Eq. (20)) indicates stronger heterogeneity across clients.

Evaluation Metrics

We evaluate all three models with two metrics, one for accuracy against the FEM reference and one for consistency with the governing equation. The global Fed-PIGANO model θ∗\theta^* and the Centralized model θcen\theta^{\mathrm{cen}} are tested on the test set Dcte\mathcal{D}_c^{\mathrm{te}} of every client (G=50G = 50 geometries, § 4.2), and each Local-only model θcloc\theta_c^{\mathrm{loc}} on its own client’s test set. For a test geometry gg, let TgFEM(x,y)T_g^{\mathrm{FEM}}(x,y) be the FEM reference temperature at a point (x,y)∈Xg(x,y) \in\mathbf{X}^g. The relative ℓ2\ell_2 error on gg, which is the same in dimensional and nondimensional units, is

ϵrel=∑(x,y)∈Xg[Tθ(x,y;Xgb)−TgFEM(x,y)]2∑(x,y)∈Xg[TgFEM(x,y)]2(22)\epsilon_{\mathrm{rel}} = \sqrt{\frac{\sum_{(x,y)\in\mathbf{X}^g}\left[T_\theta(x,y;\mathbf{X}_g^b)-T_g^{\mathrm{FEM}}(x,y)\right]^2}{\sum_{(x,y)\in\mathbf{X}^g}\left[T_g^{\mathrm{FEM}}(x,y)\right]^2}} \tag*{(22)}

The PDE-residual RMSE on gg is

ϵPDE=1Npts∑(x,y)∈Xg[Rθ(x,y;Xbg)]2(23)\epsilon_{\mathrm{PDE}} = \sqrt{\frac{1}{N_{\mathrm{pts}}} \sum_{(x,y)\in\mathcal{X}^{g}} \left[R_{\theta}(x,y;\mathbf{X}_{b}^{g})\right]^2} \tag*{(23)}

where RθR_{\theta} is the nondimensional residual of Eq. (18) and θ\theta is the evaluated model. Lower values are better for both metrics.

Results and Discussion

We compare Fed-PIGANO with the Local-only baseline at three levels of detail, with Centralized as a pooled-training reference. § 5.1 reports the mean performance over seven replications of the experiment and how it changes with the heterogeneity δ\delta. § 5.2 examines one of these replications on individual test geometries, and § 5.3 shows predicted thermal fields from the same replication.

Mean Prediction Accuracy on Unseen Test Geometries over Seven Replications

Table 5 reports the two metrics of § 4.5 for every scenario and client. The experiment is replicated seven times (§ 4.4), and each replication draws a new random split of every client’s geometries into training, validation, and test sets. Within a replication, each metric is averaged over the G=50G = 50 test geometries of a client. The table gives the mean and standard deviation of these averages over the seven replications. In S1, where all clients have identical geometry-family proportions, Local-only exhibits approximately 2.3 times the relative error and 1.9–2.0 times the PDE-residual RMSE of Fed-PIGANO. This improvement is consistent with the additional geometric diversity available through federation, as Fed-PIGANO learns from 150 training geometries compared with 50 at each individual site. The advantage persists across S2–S4, with Fed-PIGANO achieving lower mean values for both metrics in all 12 scenario–client cases. Centralized pools the geometries of all clients, which the setting of this study does not allow (§ 4.3), so Table 5 includes it only as a reference for pooled training.

We next examine how the performance on unseen geometries changes as the family compositions of the clients diverge. Figure 4 plots the means of Table 5 against δ\delta. Within a replication, each client keeps the same test set in all four scenarios, so any change along δ\delta comes only from the training and validation sets. In every scenario and for both metrics, Fed-PIGANO stays below Local-only, and their error bars do not overlap. Both Local-only metrics grow with δ\delta for every client. The growth is largest for Client 3, which trains only on the concave families in S4. Its ϵrel\epsilon_{\mathrm{rel}} rises from 4.34% in S1 to 20.46% in S4, and its ϵPDE\epsilon_{\mathrm{PDE}} from 6.43×10−26.43 \times10^{-2} to 48.76×10−248.76 \times10^{-2}. For Client 1, trained only on Q and P in S4, ϵrel\epsilon_{\mathrm{rel}} rises from 4.36% to 10.16%. In contrast, the mean ϵrel\epsilon_{\mathrm{rel}} of Fed-PIGANO stays between 1.83% and 2.04% in all 12 cases and rises by at most 0.17 percentage points from S1 to S4. Its mean ϵPDE\epsilon_{\mathrm{PDE}} stays between 3.34×10−23.34 \times10^{-2} and 3.63×10−23.63 \times10^{-2}. Its accuracy (ϵrel\epsilon_{\mathrm{rel}}) is also comparable to that of the pooled Centralized reference at every δ\delta (Table 5). Through federation, each

Table 5: Relative ℓ2\ell_2 error and PDE-residual RMSE on unseen test geometries for each scenario and client, as mean ±\pm standard deviation over seven replications. Bold marks the lower mean of Local-only and Fed-PIGANO.

MethodClient 1 ϵrel↓\epsilon_{\mathrm{rel}} \downarrow (%)Client 1 ϵPDE↓\epsilon_{\mathrm{PDE}} \downarrow (×10−2\times10^{-2})Client 2 ϵrel↓\epsilon_{\mathrm{rel}} \downarrow (%)Client 2 ϵPDE↓\epsilon_{\mathrm{PDE}} \downarrow (×10−2\times10^{-2})Client 3 ϵrel↓\epsilon_{\mathrm{rel}} \downarrow (%)Client 3 ϵPDE↓\epsilon_{\mathrm{PDE}} \downarrow (×10−2\times10^{-2})
S1 (δ=0.000\delta= 0.000)
Local-only4.36±0.414.36 \pm0.416.55±0.776.55 \pm0.774.25±0.534.25 \pm0.536.88±0.656.88 \pm0.654.34±0.404.34 \pm0.406.43±0.656.43 \pm0.65
Fed-PIGANO1.91±0.171.91 \pm0.173.34±0.253.34 \pm0.251.83±0.151.83 \pm0.153.40±0.353.40 \pm0.351.89±0.221.89 \pm0.223.45±0.283.45 \pm0.28
Centralized†1.91±0.101.91 \pm0.102.77±0.152.77 \pm0.151.82±0.101.82 \pm0.102.79±0.192.79 \pm0.191.89±0.151.89 \pm0.152.87±0.212.87 \pm0.21
S2 (δ=0.198\delta= 0.198)
Local-only4.73±0.334.73 \pm0.337.86±0.717.86 \pm0.714.30±0.454.30 \pm0.457.01±0.547.01 \pm0.544.80±0.514.80 \pm0.517.33±0.467.33 \pm0.46
Fed-PIGANO1.83±0.081.83 \pm0.083.37±0.203.37 \pm0.201.90±0.131.90 \pm0.133.35±0.173.35 \pm0.171.92±0.151.92 \pm0.153.45±0.383.45 \pm0.38
Centralized†1.80±0.111.80 \pm0.112.81±0.212.81 \pm0.211.85±0.051.85 \pm0.052.76±0.102.76 \pm0.101.87±0.121.87 \pm0.122.87±0.232.87 \pm0.23
S3 (δ=0.449\delta= 0.449)
Local-only7.55±0.817.55 \pm0.8113.64±4.5213.64 \pm4.524.41±0.484.41 \pm0.487.09±0.947.09 \pm0.946.42±0.826.42 \pm0.829.20±0.979.20 \pm0.97
Fed-PIGANO1.90±0.151.90 \pm0.153.47±0.273.47 \pm0.271.98±0.211.98 \pm0.213.59±0.203.59 \pm0.202.03±0.122.03 \pm0.123.63±0.373.63 \pm0.37
Centralized†1.86±0.121.86 \pm0.122.77±0.212.77 \pm0.211.91±0.141.91 \pm0.142.78±0.192.78 \pm0.191.90±0.111.90 \pm0.112.85±0.232.85 \pm0.23
S4 (δ=0.768\delta= 0.768)
Local-only10.16±0.8210.16 \pm0.8216.94±3.4116.94 \pm3.414.96±0.244.96 \pm0.248.08±0.868.08 \pm0.8620.46±2.0420.46 \pm2.0448.76±6.6948.76 \pm6.69
Fed-PIGANO1.99±0.161.99 \pm0.163.59±0.393.59 \pm0.392.00±0.222.00 \pm0.223.58±0.283.58 \pm0.282.04±0.202.04 \pm0.203.63±0.263.63 \pm0.26
Centralized†1.88±0.051.88 \pm0.052.82±0.152.82 \pm0.151.89±0.141.89 \pm0.142.77±0.182.77 \pm0.181.96±0.141.96 \pm0.142.92±0.312.92 \pm0.31

Table 5. ↓\downarrow Lower is better. †Reference trained on the pooled geometries of all clients, shown in italics and not part of the bold comparison.

Mean test-set errors versus heterogeneity for each client

Figure 4. Mean test-set ϵrel\epsilon_{\mathrm{rel}} (top) and ϵPDE\epsilon_{\mathrm{PDE}} (bottom) versus δ\delta for each client (columns). Error bars show ±1\pm1 standard deviation over the seven replications.

client benefits from the geometries held by the others without sharing its own. Therefore, its performance on unseen geometries is better than with local training alone and changes little as the family compositions of the clients diverge.

The results above are averages over the seven replications and over the 50 test geometries of each client. They do not show how the two metrics vary from one geometry to another. Each replication uses different test geometries, so such per-geometry values can be shown for only one replication at a time. § 5.2 examines one replication in detail.

Prediction Accuracy on Individual Unseen Test Geometries

We select one replication in which Fed-PIGANO performs slightly worse than its seven-replication average, in every scenario and for both metrics, and use it in this section and in § 5.3. The results of this replication are therefore conservative, since Fed-PIGANO performs better on average in the other six replications. For this replication, Figure 5 and Figure 6 show the mean of each metric over the 50 test geometries of each client, with ±1\pm1 standard deviation over these geometries. As in § 5.1, Fed-PIGANO has a lower mean than Local-only for every client, in every scenario, and for both metrics. Its mean ϵrel\epsilon_{\mathrm{rel}} stays between 1.79% and 2.20% in all 12 cases, whereas that of Local-only lies between 3.51% and 4.72% in S1 and reaches 21.31% for Client 3 in S4. The mean ϵPDE\epsilon_{\mathrm{PDE}} shows the same contrast. It stays between 3.29×10−23.29 \times10^{-2} and 3.93×10−23.93 \times10^{-2} for Fed-PIGANO and reaches

Mean relative error across clients for four scenarios

Figure 5. Mean ϵrel\epsilon_{\mathrm{rel}} across clients for (a) S1 to (d) S4. Error bars show ±1\pm1 standard deviation over the 50 test geometries. Lower error bars are cut at 20% of the bar height where the standard deviation exceeds 80% of the mean.

Mean PDE error across clients for four scenarios

Figure 6. Mean ϵPDE\epsilon_{\mathrm{PDE}} across clients for (a) S1 to (d) S4. Error bars as in Figure 5.

40.39×10−240.39 \times10^{-2} for Local-only. The spread over the test geometries differs in the same way. The standard deviation of the Local-only ϵrel\epsilon_{\mathrm{rel}} reaches 19.30% for Client 3 in S4, while that of Fed-PIGANO stays between 0.75% and 0.95%. Centralized is included only as the pooled-training reference (§ 4.3).

The statistics in Figure 5 and Figure 6 still summarize all 50 test geometries of a client, so they do not show whether Fed-PIGANO is better on each geometry. Figure 7 compares the two models geometry by geometry. Each marker is one test geometry, placed at its Local-only value on the horizontal axis and its Fed-PIGANO value on the vertical axis. Its shape and color give the geometry family. A marker below the dashed identity line, in the shaded region, marks a geometry on which Fed-PIGANO has the lower value, and the number in each panel counts these geometries. In every scenario–client case, Fed-PIGANO gives the lower ϵrel\epsilon_{\mathrm{rel}} on 42 to 50 of the 50 test geometries, and the lower ϵPDE\epsilon_{\mathrm{PDE}} on 37 to 50. From S1 to S4, the markers shift to the right while staying at similar heights, as the Local-only values grow and the Fed-PIGANO values do not. The largest Local-only values occur on families that are absent from, or rare in, a client’s training set. For Client 3 in S4, trained only on CP and CH, the 15 geometries with the largest Local-only ϵrel\epsilon_{\mathrm{rel}} all come from Q, P, and H. Therefore, Fed-PIGANO outperforms Local-only on most individual test geometries, not only on average, and its advantage is largest on the families that a client has not trained on. For a site, federation thus helps most on the new designs that its own data cover least.

Geometry-wise comparison plots for Scenarios 1–4

Figure 7. Geometry-wise ϵrel\epsilon_{\mathrm{rel}} (top row of each subfigure) and ϵPDE\epsilon_{\mathrm{PDE}} (bottom row) of Fed-PIGANO versus Local-only for (a) S1 to (d) S4. Markers below the identity line favor Fed-PIGANO.

Predicted Thermal Fields on Sample Test Geometries

This section shows predicted thermal fields from the replication of § 5.2. We use S4, in which the family compositions of the clients differ most (δ=0.768\delta= 0.768). For each client, we select one unseen test geometry from a family that is absent from its training set (Table 3). This mirrors a manufacturing site that receives a new geometric design not covered by its own training data. The selected geometries are CH for Clients 1 and 2, and P for Client 3. Figure 8–Figure 10 compare the FEM reference with the predicted temperature, absolute error, and absolute PDE residual of the client’s Local-only model and of Fed-PIGANO. On these geometries, Local-only gives ϵrel\epsilon_{\mathrm{rel}} values of 13.30%, 7.25%, and 17.37% at Clients 1, 2, and 3, and Fed-PIGANO gives 0.84%, 1.55%, and 1.98%.

Thermal predictions for the unseen CH geometry #45 at Client 1 in S4

Figure 8. Thermal predictions for the unseen CH geometry #45 at Client 1 in S4. The left panel shows the FEM reference. The rows show the predicted temperature, absolute error, and absolute PDE residual of each model. Error and residual color scales are shared across Figs. 8–10.

Thermal predictions for the unseen P geometry #16 at Client 3 in S4

Figure 10. Thermal predictions for the unseen P geometry #16 at Client 3 in S4. Layout as in Figure 8.

At the scale of the temperature range, the predictions of both models look similar to the FEM reference. However, the absolute error and PDE residual maps reveal a difference. On each geometry, the Fed-PIGANO error and residual stay small almost everywhere, whereas those of Local-only are much larger over wide regions. Fed-PIGANO is therefore more accurate than Local-only and satisfies the governing equation more closely. Figure 8–Figure 10 show only one geometry per client, but the per-geometry comparison in Figure 7 confirms the same advantage of Fed-PIGANO on most test geometries. This advantage is largest on families absent from a client’s training set. Thus, federation gives each client accurate predictions even on geometry families missing from its own data, without sharing any geometry.

Thermal predictions for the unseen CH geometry #49 at Client 2 in S4

Figure 9. Thermal predictions for the unseen CH geometry #49 at Client 2 in S4. Layout as in Figure 8.

Conclusion and Future Work

We addressed thermal field prediction on unseen part geometries when the geometries are distributed across AM sites that cannot share them. We proposed Fed-PIGANO, which trains a physics-informed geometry-aware model at each site and aggregates its geometry encoder and prediction network across sites. We evaluated it on 2D metal AM geometries from five families, held by three sites under four scenarios of increasing heterogeneity, over seven replications. Fed-PIGANO was more accurate than training at each site alone (Local-only) in all 12 scenario–site cases and in every replication. Even with identical family compositions, the mean relative error of Local-only was about 2.3 times that of Fed-PIGANO. As the compositions diverged, it rose up to 20.46%, while that of Fed-PIGANO stayed between 1.83% and 2.04%. The accuracy of Fed-PIGANO was also comparable to that of a centralized model trained on the pooled geometries. Future work will extend the framework to 3D geometries and transient AM thermal problems with moving heat sources, investigate coupling with microstructure and property models, and explore secure aggregation and differential privacy to further protect proprietary part designs.

Acknowledgements

During the preparation of this manuscript, the authors used large language models (LLMs) for language editing of selected portions of the text. The authors reviewed and edited all LLM-assisted content and take full responsibility for the final manuscript.

References

  1. [1]Almasri, W., Bettebghor, D., Adjed, F., Danglade, F., and Abbasa, F. (2024). Geometrically-driven generation of mechanical designs through deep convolutional GANs. Engineering Optimization, 56(1):18–35.
  2. [2]Deshmukh, K., Rien­sche, A., Bevans, B., Lane, R. J., Snyder, K., Halliday, H. S., Williams, C. B., Mirzaeifar, R., and Rao, P. (2024). Effect of processing parameters and thermal history on microstructure evolution and functional properties in laser powder bed fusion of 316L. Materials & Design, 244:113136.DOI
  3. [3]Faegh, M., Ghungrad, S., Oliveira, J. P., Rao, P., and Haghighi, A. (2025). A review on physics-informed machine learning for process-structure-property modeling in additive manufacturing. Journal of Manufacturing Processes, 133:524–555.DOI
  4. [4]Gao, H., Sun, L., and Wang, J.-X. (2021). PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving parameterized steady-state PDEs on irregular domain. Journal of Computational Physics, 428:110079.DOI
  5. [5]Gnanasambandam, R., Gotawala, N., Wu, X., Standfield, B., Dou, C., Yu, H. Z., Yu, J., and Kong, Z. (2026). Coupled physics-informed neural network for multi-physics modeling in additive friction stir deposition. IEEE Transactions on Automation Science and Engineering, 23:1631–1649.DOI
  6. [6]Gnanasambandam, R., Shen, B., Chung, J., Yue, X., and Kong, Z. (2023). Self-Scalable Tanh (Stan): Multi-scale solutions for physics-informed neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15588–15603.DOI
  7. [7]Guaglione, F., Caprio, L., Previtali, B., and Demir, A. G. (2021). Single point exposure LPBF for the production of biodegradable Zn-alloy lattice structures. Additive Manufacturing, 48:102426.DOI
  8. [8]Guo, S., Wang, W., Jiang, C., Elidrisi, M., Lee, M., Madhyastha, H. V., Al Kontar, R., and Okwudire, C. E. (2026). A collaborative process parameter recommender system for fleets of networked manufacturing machines—with application to 3D printing. Journal of Manufacturing Systems, 85:22–33.arxiv.org/abs/2506.12252
  9. [9]Johnson, C. (2009). Numerical solution of partial differential equations by the finite element method. Courier Corporation.
  10. [10]Kairouz, P. and McMahan, H. B. (2021). Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1-2):1–210.arxiv.org/abs/1912.04977
  11. [11]Kang, J., Li, G., Che, Y., Cao, X., Wan, M., Zhu, J., Luo, M., and Zhang, X. (2024). Four-dimensional hemodynamic prediction of abdominal aortic aneurysms following endovascular aneurysm repair combining physics-informed PointNet and quadratic residual networks. Physics of Fluids, 36(8).DOI
  12. [12]Kashefi, A., Guibas, L. J., and Mukerji, T. (2023). Physics-informed PointNet: On how many irregular geometries can it solve an inverse problem simultaneously? Application to linear elasticity. Journal of Machine Learning for Modeling and Computing, 4(4).arxiv.org/abs/2303.13634
  13. [13]Kashefi, A. and Mukerji, T. (2022). Physics-informed PointNet: A deep learning solver for steady-state incompressible flows and thermal fields on multiple sets of irregular geometries. Journal of Computational Physics, 468:111510.DOI
  14. [14]Kashefi, A. and Mukerji, T. (2023). Prediction of fluid flow in porous media by sparse observations and physics-informed PointNet. Neural Networks, 167:80–91.DOI
  15. [15]Kashefi, A. and Mukerji, T. (2026). Physics-Informed KAN PointNet: Deep learning for simultaneous solutions to inverse problems in incompressible flow on numerous irregular geometries. International Journal for Numerical Methods in Fluids.
  16. [16]Kingma, D. P. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.arxiv.org/abs/1412.6980
  17. [19]Liao, S., Xue, T., Jeong, J., Webster, S., Ehmann, K., and Cao, J. (2023). Hybrid thermal modeling of additive manufacturing processes using physics-informed neural networks for temperature prediction and parameter identification. Computational Mechanics, 72(3):499–512.DOI
  18. [20]Lin, J. (1991). Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory, 37(1):145–151.DOI
  19. [21]Lin, P., You, C., Wang, Q., and Hu, G. (2024). Deep learning-based fast prediction of flow field around multiple bluff bodies. Advances in Wind Engineering, page 100020.DOI
  20. [22]Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljacic, M., Hou, T., and Tegmark, M. (2025). KAN: Kolmogorov–Arnold Networks. In International Conference on Learning Representations.DOI
  21. [23]McMahan, B., Moore, E., Ramage, D., Hampson, S., and Aguera y Arcas, B. (2017). Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR.arxiv.org/abs/1602.05629
  22. [27]Raissi, M., Perdikaris, P., and Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707.DOI
  23. [28]Saimon, A. I., Gnanasambandam, R., and Kong, Z. (2026a). Geometry-Aware Physics Informed PointNet (GeoPIPN) for fast thermal distribution prediction in additive manufacturing of unseen part geometries. IEEE Transactions on Automation Science and Engineering.DOI
  24. [29]Saimon, A. I., Yangue, E., Yue, X., Kong, Z. J., and Liu, C. (2026b). Advancing additive manufacturing through deep learning: A comprehensive review of current progress and future challenges. IISE Transactions, 58(1):33–56.
  25. [30]Saimon, A. I. and Zaman, K. (2026). Likelihood-based partially decoupling approach to coupled multidisciplinary analysis under epistemic uncertainty. Research in Engineering Design, 37(1):6.DOI
  26. [31]Smith, J., Xiong, W., Yan, W., Lin, S., Cheng, P., Kafka, O. L., Wagner, G. J., Cao, J., and Liu, W. K. (2016). Linking process, structure, property, and performance for metal-based additive manufacturing: computational approaches with experimental support. Computational Mechanics, 57:583–610.DOI
  27. [33]Wang, C., Wu, J., Wang, Y., Zha, Z., and Zhou, Q. (2024). MPIPN: A Multi Physics-Informed PointNet for solving parametric acoustic-structure systems. Engineering with Computers, pages 1–22.arxiv.org/abs/2403.01132
  28. [34]Wang, R., Garcia, D., Kamath, R. R., Dou, C., Ma, X., Shen, B., Choo, H., Fezzaa, K., Yu, H. Z., and Kong, Z. (2022). In situ melt pool measurements for laser powder bed fusion using multi sensing and correlation analysis. Scientific Reports, 12(1):13716.DOI
  29. [35]Wang, R., Standfield, B., Dou, C., Law, A. C., and Kong, Z. J. (2023a). Real-time process monitoring and closed-loop control on laser power via a customized laser powder bed fusion platform. Additive Manufacturing, 66:103449.DOI
  30. [36]Wang, S., Sankaran, S., Wang, H., and Perdikaris, P. (2023b). An expert’s guide to training physics-informed neural networks. arXiv preprint arXiv:2308.08468.arxiv.org/abs/2308.08468
  31. [37]Xie, J., Chai, Z., Xu, L., Ren, X., Liu, S., and Chen, X. (2022). 3D temperature field prediction in direct energy deposition of metals using physics informed neural network. The International Journal of Advanced Manufacturing Technology, 119(5):3449–3468.
  32. [38]Zhang, H., Liu, L., Weng, K., and Lu, L. (2025a). Federated scientific machine learning for approximating functions and solving differential equations with data heterogeneity. IEEE Transactions on Neural Networks and Learning Systems, 36(10):18104–18117.arxiv.org/abs/2410.13141
  33. [39]Zhang, R., Maftouni, M., Chung, J., Shen, B., Wang, R., Law, A. C. C., Rao, P., and Kong, Z. J. J. (2025b). Automatic melt pool tracking and segmentation in laser powder bed fusion using X-ray image sequence. Journal of Intelligent Manufacturing, pages 1–17.
  34. [40]Zhong, W. and Meidani, H. (2025). Physics-Informed Geometry-Aware Neural Operator. Computer Methods in Applied Mechanics and Engineering, 434:117540.arxiv.org/abs/2408.01600
  35. [41]Zhu, Q., Liu, Z., and Yan, J. (2021). Machine learning for metal additive manufacturing: predicting temperature and melt pool fluid dynamics using physics-informed neural networks. Computational Mechanics, 67:619–635.arxiv.org/abs/2008.13547

Paper details

vlnc:W_s87934x3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

large language models (LLMs)

View statement
Looking for related papers…
Contents